krutrim-ai-sdk

Why this exists

Krutrim Cloud already has an OpenAI-compatible API. You can point the raw OpenAI client at https://cloud.olakrutrim.com/v1 and ship.

Building real products needs more:

• First-class Vercel AI SDK support — generateText, streamText, tools, structured output, Next.js streaming — not a second-class fork of openai
• Model aliases that stick — krutrim-2 → Krutrim-2, without hunting catalogue strings
• Indic-native defaults — Hindi/Tamil/Malayalam system prompts, Hinglish notes, support-agent presets, ₹/INR-aware errors
• Clear failure modes — rate limits, credits, region, bad keys — explained for Indian builders, not generic US SaaS copy

I built krutrim-ai-sdk so Krutrim works with the same DX as other AI SDK providers (similar spirit to sarvam-ai-sdk): solid defaults, typed APIs, and production-ready ergonomics.

Install (one command)

npm i krutrim-ai-sdk ai@6

pnpm add krutrim-ai-sdk ai@6
# or
yarn add krutrim-ai-sdk ai@6
# or
bun add krutrim-ai-sdk ai@6

Get an API key from the Krutrim Cloud console:

# .env — either name works
KRUTRIM_API_KEY=your_api_key_here
# KRUTRIM_CLOUD_API_KEY=your_api_key_here

Quick start

1. Basic chat

import { krutrim } from 'krutrim-ai-sdk';
import { generateText } from 'ai';

const { text } = await generateText({
  model: krutrim('Krutrim-2'),
  prompt: 'नमस्ते! मुंबई के बारे में एक मज़ेदार तथ्य बताओ।',
});

console.log(text);

2. Streaming

import { streamText } from 'ai';
import { krutrim } from 'krutrim-ai-sdk';

const result = streamText({
  model: krutrim('Krutrim-2'),
  prompt: 'Write a short poem about the Indian monsoon.',
});

for await (const delta of result.textStream) {
  process.stdout.write(delta);
}

3. Next.js App Router

app/api/chat/route.ts:

import { streamText, convertToModelMessages, type UIMessage } from 'ai';
import { krutrim, indicResponsePrompt } from 'krutrim-ai-sdk';

export const maxDuration = 60;

export async function POST(req: Request) {
  const { messages }: { messages: UIMessage[] } = await req.json();

  const result = streamText({
    model: krutrim('Krutrim-2'),
    system: indicResponsePrompt('en-IN', { allowCodeMix: true }),
    messages: await convertToModelMessages(messages),
  });

  return result.toUIMessageStreamResponse();
}

Wire any AI SDK UI (useChat) to POST /api/chat.

4. Agent with tools

import { generateText, tool, stepCountIs } from 'ai';
import { z } from 'zod';
import { krutrim, indicSupportAgentPrompt } from 'krutrim-ai-sdk';

const result = await generateText({
  model: krutrim('Krutrim-2'),
  system: indicSupportAgentPrompt({
    brandName: 'MyApp',
    languages: ['hi-IN', 'en-IN', 'ta-IN'],
  }),
  tools: {
    lookupOrder: tool({
      description: 'Look up an order by ID',
      inputSchema: z.object({ orderId: z.string() }),
      execute: async ({ orderId }) => ({
        orderId,
        status: 'Shipped',
        amountInr: 1499,
      }),
    }),
  },
  stopWhen: stepCountIs(5),
  prompt: 'Mera order ORD-9912 kahan hai?',
});

console.log(result.text);

5. Custom provider config

import { createKrutrim } from 'krutrim-ai-sdk';

const krutrim = createKrutrim({
  apiKey: process.env.KRUTRIM_API_KEY,
  baseURL: 'https://cloud.olakrutrim.com/v1',
  headers: { 'X-App-Name': 'my-app' },
});

Supported models

IDs follow the Krutrim Model Catalogue. Confirm in the console if a call returns “invalid model”.

Native Krutrim (highlight)

| Model | ID | Modality | | ----- | -- | -------- | | Krutrim-2 | Krutrim-2 / alias krutrim-2 | Text ★ default pick | | Krutrim-1 | Krutrim-1 / alias krutrim-1 | Text | | Krutrim-Dhwani | Krutrim-Dhwani | Speech-to-Text | | Krutrim-TTS | Krutrim-TTS | Text-to-Speech | | Vyakyarth | Vyakyarth | Embeddings | | Bhasantarit | Bhasantarit | Embeddings | | chitrapathak | chitrapathak | Vision / image-text |

Popular hosted open models

| Model | ID | | ----- | -- | | Llama 3.3 70B | Llama-3.3-70B-Instruct | | Llama 3.2 11B Vision | Llama-3.2-11B-Vision-Instruct | | DeepSeek R1 | DeepSeek-R1 | | Qwen3 32B | Qwen3-32B | | Mistral 7B | Mistral-7B-v0.2 | | Phi-4 Reasoning | Phi-4-reasoning-plus | | Gemma 3 27B | gemma-3-27b-it |

import { KRUTRIM_CHAT_MODELS, resolveModelId } from 'krutrim-ai-sdk';

krutrim(KRUTRIM_CHAT_MODELS.krutrim2);
resolveModelId('deepseek-r1'); // → "DeepSeek-R1"

krutrim-ai-sdk vs raw OpenAI SDK

| | OpenAI SDK pointed at Krutrim | krutrim-ai-sdk | | - | ----------------------------- | ------------------ | | Works with AI SDK generateText / streamText | Manual / wrappers | ✅ native | | Next.js UI message streaming | DIY | ✅ toUIMessageStreamResponse | | Tool calling + agents | DIY mapping | ✅ first-class | | Model aliases (krutrim-2) | ❌ | ✅ | | Indic system prompts | ❌ | ✅ indicResponsePrompt, etc. | | INR / rate-limit error tips | Generic | ✅ India-aware | | Embeddings + Bhashik helpers | Separate clients | ✅ same package | | Dependencies | openai client | Minimal (@ai-sdk/*) |

You still use Krutrim’s OpenAI-compatible endpoint under the hood — this package is the AI SDK-shaped door into it.

Indic helpers 🇮🇳

import {
  indicResponsePrompt,
  transliterationNotes,
  indicSupportAgentPrompt,
  INDIC_LANGUAGES,
} from 'krutrim-ai-sdk';

// Force native-script Hindi answers
indicResponsePrompt('hi-IN');

// Hinglish / Romanized input awareness
transliterationNotes('hi-IN');

// Support agent: mirrors language, uses ₹ when relevant
indicSupportAgentPrompt({ brandName: 'PayApp' });

Bhashik / TTT helpers (Sarvam-style surface)

Same idea as sarvam-ai-sdk: language ID, translation, transliteration, speech, STT — plus Krutrim extras.

// Language detection (also: languageIdentification())
await generateText({
  model: krutrim.languageDetection(),
  prompt: 'എന്തൊരു മനോഹരമായ ദിവസം!',
});

// Translation
await generateText({
  model: krutrim.translation({ from: 'hi-IN', to: 'en-IN' }),
  prompt: 'आज मौसम बहुत सुहाना है।',
});

// Transliteration (chat-backed; same DX as Sarvam)
await generateText({
  model: krutrim.transliterate({ to: 'hi-IN', from: 'en-IN' }),
  prompt: 'namaste, aap kaise ho?',
});

// Summarization / sentiment (Bhashik Language Labs)
await generateText({
  model: krutrim.summarization({ language: 'hin', summarySize: 40 }),
  prompt: '…long Hindi article…',
});
await generateText({
  model: krutrim.sentiment({ language: 'eng' }),
  prompt: 'The service was excellent!',
});

// Embeddings
await embed({
  model: krutrim.embedding('Vyakyarth'),
  value: 'भारत की राजधानी नई दिल्ली है।',
});

Version compatibility

| krutrim-ai-sdk | Vercel AI SDK | | -------------- | ------------- | | 0.1.x (current) | 6.x | | 0.2.x (planned) | 7.x |

Same generation as sarvam-ai-sdk 0.3.x (AI SDK v6 / LanguageModelV3).

API surface

import { krutrim } from 'krutrim-ai-sdk';

// Chat
krutrim('Krutrim-2');
krutrim.languageModel('Llama-3.3-70B-Instruct');
krutrim.chat('deepseek-r1');

// Embeddings
krutrim.embedding('Vyakyarth');
krutrim.textEmbeddingModel('Bhasantarit');

// Speech / STT (language-first or Sarvam-style model + language)
krutrim.speech('hi-IN');
krutrim.speech('Krutrim-TTS', 'hi-IN');
krutrim.transcription('ta-IN');
krutrim.transcription('Krutrim-Dhwani', 'ta-IN');

// Language ID / translation / transliterate
krutrim.languageDetection();
krutrim.languageIdentification(); // alias
krutrim.translation({ from: 'hi-IN', to: 'en-IN' });
krutrim.translation('krutrim-translate-v1.0', { from: 'hi-IN', to: 'en-IN' });
krutrim.transliterate({ to: 'hi-IN', from: 'en-IN' });

// Bhashik extras
krutrim.summarization({ language: 'hin' });
krutrim.sentiment({ language: 'eng' });

Base URL: https://cloud.olakrutrim.com/v1
Docs: Inferencing · Model catalogue

Examples

| File | Demo | | ---- | ---- | | examples/generate-text.ts | Hindi chat | | examples/stream-text.ts | Streaming | | examples/tools.ts | Tool calling | | examples/agent.ts | Multilingual agent | | examples/structured-output.ts | Structured JSON | | examples/embeddings.ts | Embeddings | | examples/nextjs-app-router/route.ts | App Router |

cp .env.example .env
npm i
npx tsx --env-file=.env examples/generate-text.ts

Roadmap

• [x] Chat completions (generate + stream)
• [x] Tool calling & structured output path
• [x] Model aliases + models.ts constants
• [x] Indic prompt helpers + India-aware errors
• [x] Embeddings (Vyakyarth / Bhasantarit)
• [x] Bhashik TTS / STT / LID / translation (lightweight)
• [x] Transliterate + languageIdentification (Sarvam-compatible surface)
• [x] Summarization + sentiment (Bhashik)
• [ ] Official listing under AI SDK community providers
• [ ] Image generation helpers (diffusion / multimodal)
• [ ] AI SDK v7 (LanguageModelV4) track
• [ ] More integration tests + sample apps (chat UI, voice)

Vote with ⭐ and Issues — popular requests get prioritized.

Contributing

PRs and issues are welcome. See CONTRIBUTING.md.

npm i
npm run build
npm test

Disclaimer

Unofficial project — not affiliated with or endorsed by Ola / Krutrim unless stated otherwise. Maintained by aljojoby9.

Links

• Krutrim Cloud
• Krutrim docs
• Vercel AI SDK
• Custom provider guide
• Sarvam AI SDK (inspiration)

License

MIT · © aljojoby9